DataClaw0: Agentic Tailoring Multimodal Data from Raw Streams

2026-06-19

AI Summary

This work addresses the inefficiency of knowledge extraction from raw multimodal data streams, which suffer from high entropy, and the limitations of existing passive annotation methods that are costly and fail to uncover deep semantic structures. To overcome these challenges, the authors propose a novel paradigm termed “agent-based data distillation,” which frames data processing as a learnable capability. This approach employs a two-stage pipeline that integrates deterministic factual anchors with generative semantic synthesis and introduces Group Relative Policy Optimization (GRPO) for policy alignment. The study establishes DataClaw0-val, the first benchmark for data distillation, and trains the DataClaw0-9B model, demonstrating significantly improved task adaptation efficiency under few-shot post-training across video generation, real-world visual question answering, and GUI navigation tasks, thereby validating its ability to produce high-information-density, task-specific data.

Abstract

Massive unstructured multimodal streams suffer from high "data entropy," impeding both efficient human knowledge acquisition and high-quality AI post-training. Existing passive annotation paradigms, heavily reliant on heuristic rules or general VLMs, are costly, monotonous, and fail to unlock the deep procedural logic embedded in raw data. We elevate data processing to a learnable capability, proposing a paradigm shift towards Agentic Data Tailoring, which actively refining and structuring data to align with diverse user and downstream intents. To overcome the data scarcity bottleneck in training such high-order capabilities, we design a two-stage pipeline grounding generative semantic synthesis in deterministic Factual Anchors, yielding a large-scale dataset spanning five core physical and digital domains. Building upon this, $\text{DataClaw}_0$-9B model synergizes Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO), achieving robust alignment with complex refinement and tailoring intents. To systematically quantify this capability, we construct $\text{DataClaw}_0$-val, the first benchmark dedicated to data refinement. Crucially, we adopt downstream post-training as the ultimate validation touchstone. Evaluations on video generation, real-world VQA, and GUI navigation confirm that $\text{DataClaw}_0$ delivers high-information-density tailored data, facilitating efficient model adaptation to new tasks under limited training data regimes. Project page: https://czjdsg.github.io/MakeAnyData

Problem

• multimodal data
• data entropy
• data tailoring
• unstructured data
• AI post-training

Innovation

• Agentic Data Tailoring
• Factual Anchors
• Group Relative Policy Optimization
• Multimodal Data Refinement
• DataClaw